- Jacob Coxon, a former pretraining researcher at OpenAI and Anthropic, has resigned over AI-safety fears.
- He estimates the odds of AI destroying humanity this decade at above ten percent.
- Coxon said the labs are “gambling with our lives,” accusing both of downplaying risk.
- His departure adds to a pattern of high-profile safety-motivated exits.
What Happened
Jacob Coxon, a former pretraining researcher at both OpenAI and Anthropic, has quit and publicly accused the companies of knowingly downplaying catastrophic risk, The Decoder reported on September 9, 2026. Speaking to Bloomberg, Coxon said the industry is “gambling with our lives” and put the odds of AI destroying humanity this decade above ten percent.
Why It Matters
Coxon is unusual in having worked on pretraining — the core model-building stage — at the two leading safety-focused labs. A person with that vantage point assigning a double-digit probability to catastrophe, and resigning over it, is a pointed internal critique rather than an outsider’s warning. It revives a debate that has produced other safety-motivated departures from frontier labs over the past two years.
Technical Details
A greater-than-ten-percent estimate is a “p(doom)” figure — a subjective probability of existential harm — not a measured quantity, and such estimates vary widely among researchers. What gives Coxon’s claim weight is not precision but provenance: he is describing his assessment from inside the pretraining teams, where capability trends are visible before public release. He argues both companies understand the risk and proceed anyway.
Who’s Affected
Anthropic and OpenAI face renewed scrutiny of the gap between their public safety commitments and their pace of development. Policymakers weighing AI regulation gain a credentialed insider voice. Employees at frontier labs see another example of a researcher concluding that internal safety processes are insufficient.
What’s Next
Neither company’s detailed response was reported at publication. Whether Coxon publishes a fuller technical account of his concerns — and whether other pretraining researchers echo them — will determine if this is an isolated resignation or the start of a broader break.
